Bone age assessment from lateral cephalograms using deep learning algorithms in the Indian population.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthodontics

Background:

• Traditional bone age assessment methods are time-consuming and prone to observer variability.
• Automated methods are needed for efficient and consistent assessment of skeletal maturity.
• This study focuses on the Indian population, addressing a gap in current research.

Purpose of the Study:

• To analyze pre-processing techniques and deep learning architectures for automated cervical vertebral maturation (CVM) assessment.
• To develop a machine learning model for determining bone age from cephalometric radiographs.
• To evaluate the efficacy of various machine learning models on the Indian population.

Main Methods:

• Utilized 383 cephalometric radiographs from individuals aged 10-36 years, labeled with CVM stages.
• Employed data augmentation techniques to address data imbalance.
• Applied pre-processing methods like Sobel filters and Canny edge detectors.
• Analyzed custom Convolutional Neural Network (CNN) architectures and pre-trained models (ResNet-50, VGG-19).

Main Results:

• Custom CNN models with 6-8 convolutional layers achieved the highest accuracy of 94% on 64x64 grayscale images.
• Pre-trained ResNet-50 (49 frozen layers) and VGG-19 (10 frozen layers) showed strong performances with 91% and 89% accuracy, respectively.
• Fastest training times were observed for models with 6 and 8 convolutional layers.

Conclusions:

• Custom deep CNN models demonstrate high accuracy for classifying cervical vertebral maturation stages.
• This research provides a foundation for developing automated bone age assessment tools from lateral cephalograms.
• The findings support the potential for machine learning in clinical applications for skeletal maturity evaluation.